Pace the Frontier: What AI Labs Actually Committed To

On 12 September 2026 Anthropic CEO Dario Amodei published an essay arguing that the AI industry should deliberately slow down, and within hours the heads of OpenAI, xAI and Google DeepMind said they agreed. The essay is called We Must Pace the Frontier, and its argument is blunt: "We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast, and we must make wise use of the time we gain." For an industry that has spent four years competing on release speed, four rivals agreeing on the brakes in a single afternoon is unusual.

The harder question is what any of them actually committed to. Agreement in an essay is not a binding promise. A reply on X is not a policy. And one of the four has been making this argument since 2023.

So this article separates the commitments from the statements. It lays out the timeline, which started six days before the essay, and it covers the part almost nobody reported: a coordinated slowdown may not be legal in the United States.

The Key Takeaways

• One commitment made, one promised: Anthropic is unilaterally giving third-party evaluators permanent, employee-level access. Sam Altman says OpenAI "will do the same" and will share more soon.
• Pacing is not pausing: Amodei writes, "To be clear, pacing does not mean halting model training or technical progress." Almost every headline dropped that sentence.
• It did not start with Amodei: OpenAI chief scientist Jakub Pachocki called for coordinated slowdowns on 6 September, six days earlier.
• The legal catch: WIRED reported that OpenAI asked members of Congress whether an industry-wide slowdown would breach the Sherman Antitrust Act.
• Safety reached the balance sheet: the same day, Altman ruled out an OpenAI listing this year, calling it an "ill-advised moment to go public". Anthropic, which wrote the essay, is reported to be marketing its own IPO from mid-October.
• Not the same as "AI is slowing down": that phrase usually means capability progress is plateauing. This is the opposite claim, that progress is too fast.

What Pace the Frontier Actually Says

The essay is published on Amodei's personal site rather than Anthropic's newsroom, which matters: it is framed as a personal argument, not a corporate announcement. You can read the full essay on Amodei's own site, and it is worth doing, because the coverage has compressed it into a slogan.

The three steps

The plan has three parts, and they escalate. First, "Each frontier AI company commits to giving ongoing, employee-like access to a team of embedded third-party evaluators." Second, "Frontier AI companies within democratic countries coordinate to establish common safety standards as well as limits on the rate of unchecked AI progress." Third, "The US and other democratic governments attempt to coordinate with authoritarian governments, to the extent this is possible", on narrower questions such as banning AI use in bioweapons development or capping the rate of recursive self-improvement.

Read in order, the three steps move from something a single company can do tomorrow to something that requires an international agreement. Only the first is available to any lab acting alone, which is exactly why it is the only one anybody has actually done.

What Anthropic committed to on its own

Anthropic, the essay says, "is unilaterally committing to this step now". In his own post announcing the essay, Amodei described it precisely: the company will "provide third-party evaluators with permanent, employee-level access to our systems, so that they can verify adherence to our safety measures, report on incidents, and assess models' alignment during training." The essay is more concrete still: desks, access badges, company laptops, the same permissions as the internal risk team, and the right to publish findings without editorial control by the company.

That last detail is the one to watch. Access without publication rights is an audit the company controls. Publication rights are what turn an embedded evaluator into something closer to an independent inspector, and they are the reason this commitment is more than a press release. Whether the other labs match that specific provision, rather than the general idea, is the thing to check in three months.

The sentence everyone skipped

Amodei is explicit that he is not proposing a pause: "To be clear, pacing does not mean halting model training or technical progress, but ensuring companies take adequate time to align and safeguard their models." He also repeats that he believes AI can enormously improve human life. Headlines that read "AI leaders call for a halt" are describing a different proposal from the one that was published.

The essay's sharpest passage is a conditional, not a forecast, and it deserves care. Amodei writes that a swarm of AI agents could, in 6 to 12 months, be capable of taking over the entire internet with a persistent botnet. That is a scenario offered to justify urgency. It is not a dated prediction that this will happen, and reporting it as one misrepresents the argument. For context on how his timelines have run before, see our profile of Anthropic's co-founder and chief executive.

Who Agreed, and What They Committed To

This is where the reporting and the reality diverge most. "Altman and Musk agree" is accurate as a summary of sentiment and useless as a description of what changed. The four positions are not equivalent.

Sam Altman and OpenAI

Altman went further than a nod. His post on 12 September reads in full: "I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks. Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We'll have more to share soon." That is an explicit commitment to match step one, not merely applause.

The caveat is the last sentence. "We'll have more to share soon" means the terms have not been published, and the terms are where this gets decided. Anthropic named permanent access, internal-risk-team permissions and publication without editorial control. Until OpenAI names its own, the correct description is a promise rather than a matching commitment. The relationship between these two labs is its own long story, which we cover in our comparison of Anthropic and OpenAI.

Elon Musk and xAI

Musk's contribution was three words: "Dario is right." It is being reported as a surprising alignment between rivals. It is not. Musk signed the Future of Life Institute's March 2023 open letter, which called for a six-month pause on training systems more powerful than GPT-4. His agreeing that AI is moving too fast is continuity, not a reversal, and xAI has announced no change to how it develops or releases Grok.

Demis Hassabis and Google DeepMind

Hassabis is reported to have backed the direction while saying the specifics need further work. What is firmly on the record is his own proposal, published on 14 July 2026 in an essay titled "A Framework for Frontier AI and the Dawning of a New Age". As reported at the time by Axios, TechCrunch and Quartz, he argued for a US-led standards body modelled on FINRA, the industry-funded regulator that polices Wall Street under government oversight. Labs would voluntarily share frontier models up to 30 days before release, for testing of dangerous cyber, biological and deception capabilities. Formalisation would follow once the regime proved effective. Our profile of the DeepMind co-founder and chief executive covers how he arrived at that position.

The IPO That Is Not Happening, and the One That Is

The costliest thing OpenAI did that day was not the post about evaluators. In an interview with Fortune at the company's San Francisco headquarters on 12 September 2026, Altman ruled out taking OpenAI public this year: "I actually think that, given everything happening with safety, right now would be an ill-advised moment to go public." Asked whether that meant no listing in 2026, he answered "I would say not 2026. Yeah, we got a lot of stuff to do." The condition he named for going public was the business and society being ready, not the market window being open.

In the same interview he was asked about a 10 percent chance that AI causes human extinction. His answer was that the exact number does not matter: "Whether it's 10 or eight or six, the point is, we all have a tremendous amount of responsibility, and cannot let egos or incentives for profit or anything else get in the way." For a company whose listing has been treated as a question of when rather than whether, that is an expensive sentence. It is also the only pacing-adjacent decision any of the four labs has taken that gives something up rather than promising a process, which is why it belongs in this story even though it is not part of the three-step plan. Our profile of the OpenAI chief executive covers how he arrived here.

The awkward half sits at the other end. Reuters reported in early September that Anthropic, the company that published the pacing essay, expects to begin marketing its own IPO from mid-October and to complete the listing days before the US midterm elections in November. Nothing about that contradicts the essay, which argues for pacing capability rather than for staying private. But it does mean the lab making the strongest case for slowing down is the one heading into public markets first, with quarterly reporting and shareholder expectations attached, and that is worth holding on to when step two comes up for negotiation.

It Started Six Days Earlier

Treating the 12 September essay as the beginning of this story misses the setup. On 6 September 2026, OpenAI chief scientist Jakub Pachocki published an essay called An Alien Mind making a strikingly similar argument from inside the company with the most to lose from slowing down.

Pachocki wrote that "no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer", and that he expects and hopes "for voluntary slowdowns to become commonplace until shared safety bars are established". He went further than Amodei on enforcement, arguing that existing frameworks should evolve into "widely mandated safety bars for continued development" that "can be enforced by a network of third-party auditors, by government agencies or by international bodies". He also expects current progress to extend into recursive self-improvement, with models contributing directly to the research that builds their successors.

Two chief-scientist-and-chief-executive pairs at rival labs publishing the same conclusion within a week is the actual news. It suggests a conversation that was already happening privately, which is where the legal problem comes from. If you want the longer arc of these predictions, our guide to when AGI might actually arrive tracks how the timelines have moved.

Why Pace the Frontier May Be Illegal in the US

This is the part almost entirely absent from the coverage, and it is the part that makes the story hard.

The Sherman Act question

WIRED reported that OpenAI approached members of the US Congress in recent weeks to establish whether orchestrating an industry-wide pause in frontier AI development would fall foul of federal antitrust law. The concern, according to that reporting, is that coordinating with rivals such as Anthropic and Google on safety-related slowdowns could be read as competitors jointly restricting output. That is the classic shape of a Sherman Antitrust Act violation. Legal experts quoted in the same report drew the line at the point where coordination stops looking like sharing safety information and starts looking like a coordinated restriction on production.

This claim rests on a single outlet's sourcing and should be read as reporting rather than as established fact. But it explains the shape of Amodei's proposal. Step one is unilateral precisely because a company can do it alone without agreeing anything with a competitor. Step two, coordinating on common standards and limits, is the step that runs straight into the statute.

The bill that would change it, and what it actually covers

There is legislation in play, and it is routinely described more broadly than it deserves. The Collaboration on Adversarial Threats and Security Risks Act was introduced in July 2026 in both chambers of the 119th Congress. It is S.5105 in the Senate, sponsored by Adam Schiff with Jim Banks, and H.R.9914 in the House, sponsored by Bob Latta with eight cosponsors. Its stated purpose is to "establish the applicability of antitrust laws to the sharing of artificial intelligence frontier model risks".

It is worth reading the bill's own framing rather than the commentary. Its sponsors present it as a narrow antitrust exemption letting American AI firms coordinate against security threats from Chinese competitors, including model distillation, not as a general licence for the industry to agree on a slowdown. It has not been enacted. So the legal route to step two currently does not exist, and the bill most often cited as the fix was written for a different problem.

The Case Against Slowing Down

The four-way agreement makes this look settled. It is not, and the objections are substantive rather than merely commercial.

The strongest is verification. Even if American labs paced themselves, there is no reliable way to confirm Chinese labs were doing the same. Monitoring compute at that scale is a hard technical problem, not a paperwork one. A second objection is that a formal slowdown politicises frontier research and hands an advantage to whoever declines to participate. A third comes from the open-source community, which argues that the cost of complying with a safety regime becomes a barrier to entry, entrenching the handful of large firms that can afford it. That last one bites hardest on exactly the open-weight labs that have been closing the capability gap.

There is also a naming problem worth clearing up, because it produces genuine confusion. "AI is slowing down" is a separate and much older argument, which holds that capability gains are plateauing and that the investment around them looks like a bubble. That claim is about whether progress can continue at this rate. Amodei's is the opposite: that it can, and should not. If you are reading about an AI slowdown, check which of the two is meant. For a record of what has actually gone wrong rather than what might, our running list of AI safety incidents in 2026 tracks the real cases.

What This Means If You Just Use AI Every Day

Practically, very little changes this month. Nobody has announced a delayed model, a slower release cadence or a capped capability. The one concrete change is that outside evaluators will be sitting inside Anthropic with real access, which should over time produce more public information about how frontier models behave before release rather than after.

The useful habit is to watch for the specific thing rather than the sentiment. If OpenAI, Google DeepMind or xAI publish an evaluator arrangement with the detail Anthropic gave, including publication rights, this becomes an industry norm. If they do not, then what happened on 12 September was four executives agreeing in public and one company acting. In the meantime, the practical hedge for anyone who relies on these models is not to build a workflow that depends on a single lab's release schedule or a single model's behaviour.

The Bottom Line

Four rival labs agreed in public that AI capability is outrunning AI safety, which is a real shift in what the industry is willing to say out loud. One backed it with a commitment that has teeth, a second promised to match it without yet saying how, and two offered support. The step that would actually slow the industry down, coordinating on shared limits, is the one that may require an act of Congress before anyone can legally attempt it. Judge this by what gets published next, not by what was posted on 12 September.

FAQ